liyucs / ragnet Goto Github PK

I don't have a gt image. How to get prediction output on a single image? What these real20 and real45 names come from? I followed this discussion but it didn't help me. Would you please walk us through this? And please add this in your read-me file too.

Dear author,

Thanks in advance for your great work and effort!

I did some tests with my own images, and had some problems. The main problem is about the result of de-reflection images.

As you can see, some parts of the images have obvious purple dots, I am wondering whether these errors are caused by my side or they are the normal outputs.

If the answer is the latter, could you please explain why these purple dots are generated and are you going to reduce this side effects in the future? If the answer is the former, could you please inform me of the correct method to use your model? What I did is to replace my images with the images in folder ./testsets/real45 and run python test.py

Samples in my test are showed below:

Thanks for sharing your work. I'm interesting RAGNet is good at removing strong reflection. So I have trained again.
But the result is different with your pre-trained model(result images and sir2 score).
I used Pascal VOC dataset 2012(reflection: 7643 images, transmission: 7643 images, randomly be chosen) to synthetic dataset and Berkeley real dataset(90 images) to real dataset.
I didn't crop the images since they are cropped in your code.
Is there any mistake? Should I have to crop the image before using?
Can you explain the status of dataset?
Thank you in advance.

Hi, thank you for sharing this amazing work!

Can you explain what is the purpose of 'Fusion Ratio" while preparing dataset in the code shared?

Many thanks for your work. But I am confused about the dataset used in your codes. For 90 real-world training images from Berkeley real dataset, there are only the 'blended' and 'transmission_layer' images while for the synthetic dataset, the input is the trasmission and the reflection images. But in your 'prepare_data', the input is the transmission_layer, reflection_layer and blended images. So I am sorry to ask would you please give more details about the dataset construction. Many thanks.